Predictive and Explainable Machine Learning Framework for Sustainable Irrigation and Crop Water Optimization

Agricultural irrigation is critical in maintaining food production in the global arena yet excessive use of water is a huge challenge due to population pressures and scarce water resources. Conventional irrigation methods are usually very wasteful of water and they do not provide you with the precision in the utilization of the resources. In this document, two datasets are utilized, in order to address these problems: a free irrigation dataset on GitHub and a dataset which was created out of the base features. The datasets were highly preprocessed with many operations being done to them which included removing null values, duplicates, adding more features, normalization by MinMax Scaler, and the use of SMOTe with StandardScaler in both the LSTM and stacking classifier models in best-time classification. The tasks of classification such as irrigation and best-time classification were performed using ANN with TanELU, ReLU, Tanh, and ELU activation functions, LSTM networks, and stacking classifiers. Forecasting regression tasks were done using random forest, XGBoost, ADABoost, stacking regressor, and voting regressor. Accuracy, precision, recall, F1-score and training time were used as metrics to classify. Measures to be used in regression were MSE, RMSE and MAE. Stacking classifier performed better than the other two and obtained 99.9% best time classification, 99.6% irrigation classification and 0.288 RMSE predicting regression. The most important features were demonstrated by explainable AI methods, including LIME and SHAP. Real-time input-based prediction became possible with a Flask-based user interface, which demonstrated a tremendous difference in controlling less-water-intensive agricultural irrigation.